Large model-based auxiliary mediation method and system, and storage medium

Abstract

The application provides a mediation method and system based on a large model, and a storage medium. The method comprises the following steps: obtaining dispute exchange speech, and performing speaker analysis on the dispute exchange speech to obtain a speaker analysis result; performing speaker segmentation on the dispute exchange speech according to the speaker analysis result to obtain speaker speech, and performing speech recognition on the speaker speech to obtain dispute speaker text; inputting the dispute speaker text into a pre-trained large model to perform mediation analysis, obtaining mediation suggestions and a risk level, and sending mediation prompts to mediators according to the mediation suggestions; and if the risk level is greater than or equal to a level threshold, performing dispute risk control alarm on the mediators. In the embodiment of the application, mediation suggestions are automatically generated in a large model-based manner, mediation prompts are sent to mediators through the mediation suggestions, the decision-making cost of the mediators is reduced, mediation achieves results more quickly, and the mediation efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an auxiliary mediation method and system based on a large model and a storage medium. BACKGROUND

[0002] In the public security and judicial system, mediation is an effective means to solve the contradictions and disputes of the people in the jurisdiction and to eliminate risks in the subtle place. In order to improve the utilization rate of judicial resources, the problem of mediation efficiency is increasingly valued by people.

[0003] In the existing mediation process, mediation is generally carried out in a manual manner based on mediators, resulting in a serious shortage of mediators, the quality of mediators cannot be guaranteed, and the mediation efficiency is reduced. SUMMARY

[0004] The purpose of the embodiment of the present application is to provide an auxiliary mediation method and system based on a large model and a storage medium to solve the problem of low mediation efficiency in the prior art.

[0005] The embodiment of the present application is implemented in the following manner: an auxiliary mediation method based on a large model, the method comprising:

[0006] obtaining dispute exchange speech and performing speaker analysis on the dispute exchange speech to obtain a speaker analysis result;

[0007] performing speaker segmentation on the dispute exchange speech according to the speaker analysis result to obtain speaker speech, and performing speech recognition on the speaker speech to obtain dispute speaker text;

[0008] inputting the dispute speaker text into a pre-trained large model to perform mediation analysis, obtaining mediation suggestions and a risk level, and sending mediation prompts to mediators according to the mediation suggestions;

[0009] if the risk level is greater than or equal to a level threshold, performing a dispute risk control alarm to the mediator.

[0010] Preferably, before inputting the dispute speaker text into the pre-trained large model to perform mediation analysis, the method further comprises:

[0011] obtaining a dispute speaker sample and inputting the dispute speaker sample into the large model to perform word segmentation to obtain a sample word segmentation;

[0012] performing feature extraction on the sample word segmentation to obtain context sample features, and performing feature decoding on the context sample features to obtain decoded sample features;

[0013] mediation prediction is performed according to the decoded sample features, to obtain a sample prediction suggestion and a sample prediction level, and a model loss is determined according to the sample prediction suggestion and the sample prediction level;

[0014] The large model is updated in parameters according to the model loss until the large model converges, to obtain the pre-trained large model.

[0015] Preferably, feature extraction is performed on the sample word segmentation to obtain context sample features, including:

[0016] Word embedding processing is performed on the sample word segmentation to obtain a word vector, and self-attention mechanism calculation is performed on the word vector to obtain a self-attention vector;

[0017] Multi-head attention mechanism calculation is performed on the self-attention vector to obtain a multi-head attention vector, and normalization processing is performed on the multi-head attention vector to obtain normalized features;

[0018] Pre-feedback processing is performed on the normalized features to obtain pre-feedback features, and residual processing is performed on the pre-feedback features to obtain the context sample features.

[0019] Preferably, speaker analysis is performed on the dispute exchange voice to obtain a speaker analysis result, including:

[0020] Speech segmentation is performed on the dispute exchange voice to obtain segmented exchange voice, and a mel-frequency cepstrum coefficient, a fundamental frequency and a formant frequency of the segmented exchange voice are obtained;

[0021] The mel-frequency cepstrum coefficient, the fundamental frequency and the formant frequency are combined to obtain segmented voice features, and vector conversion is performed on the segmented voice features to obtain segmented feature vectors;

[0022] Classification is performed on the segmented exchange voice according to the segmented feature vectors to obtain a segmented voice set, and a face image of a dispute person corresponding to the dispute exchange voice is obtained;

[0023] A pronunciation time set is determined according to the face image, and a voice correspondence relationship between the dispute person and the segmented voice set is determined according to the pronunciation time set, to obtain the speaker analysis result.

[0024] Preferably, classification is performed on the segmented exchange voice according to the segmented feature vectors to obtain a segmented voice set, including:

[0025] Similarity between adjacent segmented feature vectors is calculated to obtain a first vector similarity;

[0026] If the similarity of the first vector is greater than the first similarity threshold, then the speech segments corresponding to the segmentation feature vectors are merged to obtain the segmented and merged speech.

[0027] Calculate the average vector of the segmentation feature vectors in the segmented and merged speech to obtain the segmentation average vector, and calculate the similarity between different segmentation average vectors to obtain the second vector similarity;

[0028] If the second vector similarity is greater than the second similarity threshold, then the different segmented and merged speech corresponding to the second vector similarity are divided into sets to obtain the segmented speech set.

[0029] Preferably, determining the set of pronunciation times based on the facial image includes:

[0030] The face image is subjected to feature point recognition to obtain face feature points, and the lip image is determined based on the face feature points;

[0031] Obtain the opening distance between the upper and lower lips in the lip image, and calculate the distance difference between the opening distances of adjacent lip images;

[0032] If the distance difference is greater than the difference threshold, the adjacent lip image corresponding to the distance difference is determined as the speaking image, and the image acquisition time of the speaking image is obtained to obtain the pronunciation acquisition time;

[0033] The sound acquisition time is stored in a preset set to obtain the sound acquisition time set.

[0034] Preferably, after sending a mediation prompt to the mediator based on the mediation suggestion, the method further includes:

[0035] Obtain mediation feedback on the mediation recommendation, and generate parameter fine-tuning samples based on the mediation feedback, the dispute speaker's text, and the mediation recommendation;

[0036] The parameters of the pre-trained large model are adjusted by fine-tuning the samples based on the parameters.

[0037] Another objective of this invention is to provide a large-model-based assisted mediation system, the system comprising:

[0038] The speaker analysis module is used to acquire dispute communication voice recordings and perform speaker analysis on the dispute communication voice recordings to obtain speaker analysis results;

[0039] The speech recognition module is used to segment the disputed communication speech according to the speaker analysis results to obtain the speaker's speech, and to perform speech recognition on the speaker's speech to obtain the disputed speaker's text;

[0040] The mediation prompt module is used to input the dispute speakers' text into a pre-trained large model for mediation analysis, obtain mediation suggestions and risk levels, and send mediation prompts to the mediators based on the mediation suggestions; if the risk level is greater than the level threshold, a dispute risk control alarm is sent to the mediators.

[0041] Preferably, the mediation prompt module is further used for:

[0042] Obtain a sample of the speakers in the dispute, and input the sample of the speakers in the dispute into the large model for word segmentation to obtain sample word segmentation;

[0043] The sample words are segmented and feature extraction is performed to obtain context sample features, and the context sample features are decoded to obtain decoded sample features;

[0044] Based on the features of the decoded samples, a mediation prediction is performed to obtain sample prediction suggestions and sample prediction levels, and the model loss is determined based on the sample prediction suggestions and sample prediction levels.

[0045] The parameters of the large model are updated based on the model loss until the large model converges, thus obtaining the pre-trained large model.

[0046] In this embodiment of the invention, by performing speaker analysis on the disputed communication voice, the speaker's voice corresponding to each speaker in the disputed communication voice can be effectively identified. By performing speech recognition on the speaker's voice, the speaker's voice can be effectively converted into disputed speaker text. By inputting the disputed speaker text into a pre-trained large model for mediation analysis, mediation suggestions are automatically generated in a large model-based manner. Mediation suggestions are sent to mediators to provide mediation prompts, reducing the mediator's decision-making costs, enabling mediation to achieve results more quickly, and improving mediation efficiency. Attached Figure Description

[0047] Figure 1 This is a flowchart of the large-model-based assisted mediation method provided in the first embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of the structure of the large-model-based auxiliary mediation system provided in the second embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram illustrating the specific implementation steps of the large-model-based auxiliary mediation system provided in the second embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the structure of the terminal device provided in the third embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] To illustrate the technical solution described in this invention, specific embodiments are described below.

[0053] Example 1

[0054] Please see Figure 1 This is a flowchart of a large-model-based assisted mediation method provided in the first embodiment of the present invention. This large-model-based assisted mediation method can be applied to any device or system, and includes the following steps:

[0055] Step S10: Acquire the dispute communication voice recordings and perform speaker analysis on the dispute communication voice recordings to obtain speaker analysis results;

[0056] Optionally, speaker analysis is performed on the disputed communication audio to obtain speaker analysis results, including:

[0057] The disputed speech is segmented to obtain the segmented speech, and the Mel frequency cepstral coefficients, fundamental frequency, and formant frequency of the segmented speech are obtained. Specifically, the segmented speech is passed through a Mel filter bank to obtain the energy of different Mel frequency bands, and discrete cosine transform is performed to obtain the Mel frequency cepstral coefficients. The Mel frequency cepstral coefficients can effectively reflect the spectral envelope characteristics of the speech. The fundamental frequency reflects the frequency of vocal cord vibration, and the fundamental frequency range often differs among different speakers. The formant frequency is related to the shape and size of the vocal tract and also has speaker specificity.

[0058] The Mel frequency cepstral coefficients, the fundamental frequency, and the formant frequency are combined to obtain segmented speech features, and the segmented speech features are then vectorized to obtain segmented feature vectors. In particular, by combining the Mel frequency cepstral coefficients, the fundamental frequency, and the formant frequency, segmented speech features that reflect the speaker's speech representation can be effectively obtained.

[0059] The communication segmented speech is classified according to the segmentation feature vector to obtain a segmented speech set, and the facial images of the disputants corresponding to the disputed communication speech are obtained.

[0060] The speech time set is determined based on the facial image, and the speech correspondence between the person involved in the dispute and the segmented speech set is determined based on the speech time set, thus obtaining the speaker analysis result; wherein, the speech correspondence between the person involved in the dispute and the segmented speech set is determined based on the time matching relationship between the speech acquisition time in the speech time set and the communication segmented speech in the segmented speech set.

[0061] Furthermore, the segmented speech is classified based on the segmentation feature vector to obtain a segmented speech set, including:

[0062] Calculate the similarity between adjacent segmentation feature vectors to obtain a first vector similarity; wherein, the distance between adjacent segmentation feature vectors can be calculated using the Euclidean distance formula to obtain the first vector similarity;

[0063] If the similarity of the first vector is greater than the first similarity threshold, then the speech segments corresponding to the segmentation feature vector are merged to obtain the merged speech segments. The first similarity threshold can be set according to the requirements. If the similarity of the first vector is greater than the first similarity threshold, then it is determined that the speech segments corresponding to the segmentation feature vector are spoken by the same speaker. Therefore, the speech segments corresponding to the segmentation feature vector are merged to obtain the merged speech segments.

[0064] Calculate the average vector of the segmentation feature vectors in the segmented and merged speech to obtain the segmentation average vector, and calculate the similarity between different segmentation average vectors to obtain the second vector similarity;

[0065] If the second vector similarity is greater than the second similarity threshold, then the different segmented and merged speech corresponding to the second vector similarity are divided into sets to obtain the segmented speech set; wherein, the second similarity threshold can be set according to the requirements. If the second vector similarity is greater than the second similarity threshold, it is determined that the different segmented and merged speech corresponding to the second vector similarity are speech produced by the same person. Therefore, the different segmented and merged speech corresponding to the second vector similarity are divided into sets to obtain the segmented speech set.

[0066] Furthermore, the set of pronunciation times is determined based on the facial image, including:

[0067] The face image is subjected to feature point recognition to obtain face feature points, and the lip image is determined based on the face feature points; wherein, by performing feature point recognition on the face image, face feature points in the face image can be effectively extracted, the lip position is determined based on the face feature points, and the lip image is determined based on the lip position.

[0068] Obtain the opening distance between the upper and lower lips in the lip image, and calculate the distance difference between the opening distances of adjacent lip images;

[0069] If the distance difference is greater than the difference threshold, the adjacent lip image corresponding to the distance difference is determined as the speaking image. The image acquisition time of the speaking image is obtained to obtain the pronunciation acquisition time. The pronunciation acquisition time is stored in a preset set to obtain the pronunciation time set. The difference threshold can be set according to the requirements. If the distance difference is greater than the difference threshold, it is determined that the speaker corresponding to the adjacent lip image is speaking. Therefore, the image acquisition time of the speaking image is obtained to obtain the pronunciation acquisition time.

[0070] Step S20: Based on the speaker analysis results, perform speaker segmentation on the dispute communication voice to obtain speaker voice, and perform speech recognition on the speaker voice to obtain the dispute speaker text;

[0071] Specifically, based on the speaker analysis results, the dispute communication voice is distinguished by speaker to obtain the speaker's voice. Then, by performing speech recognition on the speaker's voice, the dispute speaker's text corresponding to each speaker is obtained.

[0072] Step S30: Input the dispute speaker's text into the pre-trained large model for mediation analysis to obtain mediation suggestions and risk levels, and send mediation prompts to the mediators based on the mediation suggestions;

[0073] Among these methods, by inputting the dispute speakers' text into a pre-trained large model for mediation analysis, mediation suggestions are automatically generated based on the large model. These suggestions are then sent to mediators, reducing their decision-making costs.

[0074] Optionally, before inputting the dispute speakers' text into the pre-trained large model for mediation analysis, the following steps are also included:

[0075] The process involves obtaining dispute speaker samples and inputting them into the large-scale model for word segmentation to obtain sample word segments. This includes collecting mediation-related data from various channels, such as real mediation case documents, legal mediation document databases, and case sets from professional mediation institutions, ensuring the corpus covers multiple types of mediation cases, such as civil disputes, labor disputes, and neighborhood conflicts. The collected corpus is then cleaned, removing duplicates, errors, and non-standard formatting. For example, typos are corrected, missing information is added, and date formats are standardized. The corpus is annotated to clarify different parts of the information, such as basic case details, mediation process, and mediation results. For risk assessment-related information, key factors affecting risk assessment and their corresponding risk levels are annotated, resulting in dispute speaker samples. Based on factors such as computing resources and model performance, suitable pre-trained large-scale models, such as the GPT series and BERT, are evaluated and selected. The model's versatility in natural language processing tasks, parameter count, and adaptability to fine-tuning are considered.

[0076] The sample word segmentation is used to extract features to obtain context sample features, and the context sample features are then decoded to obtain decoded sample features; wherein, by extracting features from the sample word segmentation, context sample features containing context representation are obtained;

[0077] Mediation predictions are performed based on the decoded sample features to obtain sample prediction suggestions and sample prediction levels. The model loss is then determined based on these suggestions and levels. Specifically, based on the decoder in the large model, mediation suggestions are output through word-by-word generation. In this step, one or more fully connected layers map the decoded sample features to the risk assessment dimension. For example, if risks are divided into high, medium, and low levels, the large model will ultimately output a three-dimensional vector, where each dimension represents the probability of the corresponding risk level, thus obtaining the sample prediction level. Cross-entropy loss is typically used to measure the difference between the generated sample prediction suggestions and the actual mediation suggestions. For each generated term, the large model predicts a term probability distribution. Cross-entropy loss calculates the difference between the predicted distribution and the actual term (i.e., the term in the correctly labeled mediation suggestion). By minimizing cross-entropy loss, the model gradually adjusts its parameters to make the generated mediation suggestions closer to reality. For risk assessment, commonly used loss functions are cross-entropy loss or mean squared error loss. Taking classification problems as an example, the cross-entropy loss between the model's output risk level probability distribution and the actual labeled risk level guides the large model to adjust its parameters, improving the accuracy of risk assessment.

[0078] The parameters of the large model are updated based on the model loss until the large model converges, resulting in the pre-trained large model. After calculating the model loss, the gradient of the loss with respect to the parameters of the large model is calculated using the backpropagation algorithm. The gradient indicates how a small change in the parameters will affect the loss value and indicates the direction of parameter updates. The optimizer (such as Adam, SGD, etc.) is used to update the model parameters based on the calculated gradient.

[0079] Furthermore, feature extraction is performed on the segmented words of the samples to obtain contextual sample features, including:

[0080] The sample words are segmented and word embedding is performed to obtain word vectors. The word vectors are then processed by a self-attention mechanism to obtain self-attention vectors. Each word vector is used to calculate the degree of association with other word vectors in the text through the self-attention mechanism, thereby capturing long-distance dependencies in the text. For example, when analyzing the statements of the parties in a mediation case, the large model can use the self-attention mechanism to focus on key information, such as the focus of the dispute and the demands of the parties.

[0081] The self-attention vector is calculated using a multi-head attention mechanism to obtain a multi-head attention vector, which is then normalized to obtain normalized features. This is achieved by calculating the self-attention vector using a multi-head attention mechanism based on multiple Transformer blocks in a large model. The output of the self-attention mechanism passes through multiple Transformer blocks, which in turn calculate the attention of multiple heads in parallel. Each head focuses on different aspects of the input sequence, thereby capturing richer semantic information and dependencies. Specifically, by calculating the similarity between the query, key, and value, a weighted contextual representation is generated for each position.

[0082] The normalized features are subjected to feedforward processing to obtain feedforward features, and residual processing is performed on the feedforward features to obtain the context sample features. The output of the multi-head attention mechanism undergoes layer normalization to normalize the data, accelerating large model convergence and reducing gradient vanishing or exploding problems. The normalized result is input into a feedforward neural network, typically containing two fully connected layers, for further feature extraction and transformation. After the output of the feedforward neural network, a residual connection is established, adding the input of the current layer to the output of the feedforward neural network to help the model train better, solve gradient vanishing and degradation problems, and enable the model to learn long-term dependencies more easily.

[0083] Furthermore, after sending a mediation prompt to the mediator based on the mediation suggestion, the method further includes: obtaining mediation feedback on the mediation suggestion; generating parameter fine-tuning samples based on the mediation feedback, the dispute speaker's text, and the mediation suggestion; and adjusting the parameters of the pre-trained large model based on the parameter fine-tuning samples. Specifically, the mediation feedback from the mediator on the mediation suggestion effectively generates parameter fine-tuning samples, and the parameters of the pre-trained large model are adjusted using these samples to achieve real-time continuous model optimization, thereby improving the model quality of the pre-trained large model.

[0084] Step S40: If the risk level is greater than the level threshold, a dispute risk control alarm is sent to the mediator.

[0085] The risk level threshold can be set according to needs. If the risk level is greater than the threshold, the current dispute is considered to be serious and needs to be transferred to the public security department (to handle more serious crimes). Therefore, a dispute risk control alarm is triggered to the mediator to prompt the mediator to transfer the case.

[0086] In this embodiment, by performing speaker analysis on the dispute communication voice, the speaker voice corresponding to each speaker in the dispute communication voice can be effectively identified. By performing speech recognition on the speaker voice, the speaker voice can be effectively converted into dispute speaker text. By inputting the dispute speaker text into a pre-trained large model for mediation analysis, mediation suggestions are automatically generated in a large model-based manner. Mediation suggestions are sent to the mediator, reducing the mediator's decision-making cost, enabling the mediation to achieve results more quickly, and improving mediation efficiency.

[0087] Example 2

[0088] Please see Figure 2 This is a schematic diagram of the structure of the large-model-based auxiliary mediation system 100 provided in the second embodiment of the present invention, including:

[0089] The speaker analysis module 10 is used to acquire dispute communication voice recordings and perform speaker analysis on the dispute communication voice recordings to obtain speaker analysis results.

[0090] Optionally, the speaker analysis module 10 is further configured to: perform speech segmentation on the disputed communication speech to obtain the communication segmented speech, and obtain the Mel frequency cepstral coefficients, fundamental frequency and formant frequency of the communication segmented speech;

[0091] The Mel frequency cepstral coefficients, the fundamental frequency, and the formant frequency are combined to obtain segmented speech features, and the segmented speech features are vector-transformed to obtain segmented feature vectors.

[0092] The communication segmented speech is classified according to the segmentation feature vector to obtain a segmented speech set, and the facial images of the disputants corresponding to the disputed communication speech are obtained.

[0093] Based on the facial image, a set of pronunciation times is determined, and based on the set of pronunciation times, the speech correspondence between the person involved in the dispute and the segmented speech set is determined, thus obtaining the speaker analysis result.

[0094] Furthermore, the speaker analysis module 10 is also used to: calculate the similarity between adjacent segmentation feature vectors to obtain a first vector similarity;

[0095] If the similarity of the first vector is greater than the first similarity threshold, then the speech segments corresponding to the segmentation feature vectors are merged to obtain the segmented and merged speech.

[0096] Calculate the average vector of the segmentation feature vectors in the segmented and merged speech to obtain the segmentation average vector, and calculate the similarity between different segmentation average vectors to obtain the second vector similarity;

[0097] If the second vector similarity is greater than the second similarity threshold, then the different segmented and merged speech corresponding to the second vector similarity are divided into sets to obtain the segmented speech set.

[0098] Furthermore, the speaker analysis module 10 is also used to: perform feature point recognition on the face image to obtain face feature points, and determine the lip image based on the face feature points;

[0099] Obtain the opening distance between the upper and lower lips in the lip image, and calculate the distance difference between the opening distances of adjacent lip images;

[0100] If the distance difference is greater than the difference threshold, the adjacent lip image corresponding to the distance difference is determined as the speaking image, and the image acquisition time of the speaking image is obtained to obtain the pronunciation acquisition time;

[0101] The sound acquisition time is stored in a preset set to obtain the sound acquisition time set.

[0102] The speech recognition module 11 is used to perform speaker segmentation on the dispute communication speech based on the speaker analysis results to obtain the speaker speech, and to perform speech recognition on the speaker speech to obtain the dispute speaker text.

[0103] The mediation prompt module 12 is used to input the text of the dispute speaker into a pre-trained large model for mediation analysis, obtain mediation suggestions and risk levels, and send mediation prompts to the mediator based on the mediation suggestions; if the risk level is greater than the level threshold, a dispute risk control alarm is sent to the mediator.

[0104] Optionally, the mediation prompt module 12 is also used to: obtain a sample of the dispute speaker, and input the sample of the dispute speaker into the large model for word segmentation to obtain sample word segmentation;

[0105] The sample words are segmented and feature extraction is performed to obtain context sample features, and the context sample features are decoded to obtain decoded sample features;

[0106] Based on the features of the decoded samples, a mediation prediction is performed to obtain sample prediction suggestions and sample prediction levels, and the model loss is determined based on the sample prediction suggestions and sample prediction levels.

[0107] The parameters of the large model are updated based on the model loss until the large model converges, thus obtaining the pre-trained large model.

[0108] Furthermore, the mediation prompt module 12 is also used to: perform word embedding processing on the sample word segmentation to obtain word vectors, and perform self-attention mechanism calculation on the word vectors to obtain self-attention vectors;

[0109] The self-attention vector is calculated using a multi-head attention mechanism to obtain a multi-head attention vector, and the multi-head attention vector is then normalized to obtain normalized features.

[0110] The normalized features are subjected to pre-feedback processing to obtain pre-feedback features, and the pre-feedback features are subjected to residual processing to obtain the context sample features.

[0111] Preferably, the mediation prompt module 12 is further configured to: obtain mediation feedback on the mediation suggestion, and generate parameter fine-tuning samples based on the mediation feedback, the dispute speaker's text, and the mediation suggestion;

[0112] The parameters of the pre-trained large model are adjusted by fine-tuning the samples based on the parameters.

[0113] Please see Figure 3 The specific implementation steps of the large-model-based assisted mediation system 100 include:

[0114] 1. During the mediation process, the large-scale model-based auxiliary mediation system 100 is activated to conduct real-time audio recording;

[0115] 2. It utilizes TTS speech recognition technology to convert the received audio content into text and can distinguish different speakers based on their timbre;

[0116] 3. Use mediation corpus to fine-tune the large model, so that the large model has professional knowledge in the field of mediation, can generate mediation suggestions in real time based on the mediation content, and conduct risk assessment of the case;

[0117] 4. During the mediation process, the big data model processes the dialogue content, generates mediation suggestions in real time, and assesses the risk level of the current case.

[0118] 5. Police officers can conduct more professional mediation between the parties involved in the case based on mediation suggestions; and pay close attention to the risk assessment level. When the large model assesses the case as serious, a risk control alarm will be issued. After confirmation by the police officers, the case can be transferred to the public security department in a timely manner (to handle more serious crimes).

[0119] 6. During the mediation process, the system updates mediation suggestions and risk levels in real time;

[0120] 7. Regularly use mediation data to further fine-tune the large model, thereby enhancing its professional capabilities and risk assessment capabilities in the field of mediation.

[0121] By employing systematic risk assessment and intelligent mediation suggestions, the decision-making costs for mediators are reduced, enabling mediation to achieve results more quickly and thus improving efficiency. The large-scale model generates mediation suggestions based on the dialogue context, which is more objective compared to the traditional method of mediators providing suggestions based on personal experience. Before deployment, the system undergoes effectiveness evaluation to ensure that the implemented product achieves an above-average mediation accuracy rate. Furthermore, as an aid to mediators, it allows for mutual error correction, reducing the probability of system or human errors.

[0122] In this embodiment, by performing speaker analysis on the dispute communication voice, the speaker voice corresponding to each speaker in the dispute communication voice can be effectively identified. By performing speech recognition on the speaker voice, the speaker voice can be effectively converted into dispute speaker text. By inputting the dispute speaker text into a pre-trained large model for mediation analysis, mediation suggestions are automatically generated in a large model-based manner. Mediation suggestions are sent to the mediator, reducing the mediator's decision-making cost, enabling the mediation to achieve results more quickly, and improving mediation efficiency.

[0123] Example 3

[0124] Figure 4 This is a structural block diagram of a terminal device 2 provided in the third embodiment of this application. For example... Figure 4As shown, the terminal device 2 in this embodiment includes a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program for a large-model-based assisted mediation method. When the processor 20 executes the computer program 22, it implements the steps in the various embodiments of the large-model-based assisted mediation methods described above.

[0125] For example, the computer program 22 may be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 22 in the terminal device 2. The terminal device may include, but is not limited to, the processor 20 and the memory 21.

[0126] The processor 20 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0127] The memory 21 can be an internal storage unit of the terminal device 2, such as a hard drive or memory of the terminal device 2. The memory 21 can also be an external storage device of the terminal device 2, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal device 2. Furthermore, the memory 21 can include both internal and external storage units of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 can also be used to temporarily store data that has been output or will be output.

[0128] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0129] If an integrated module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. This computer-readable storage medium can be non-volatile or volatile. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of a computer-readable storage medium may be appropriately added to or subtracted from the contents as required by the legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, a computer-readable storage medium may not include electrical carrier signals and telecommunication signals.

[0130] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A large-model-based assisted mediation method, characterized in that, The method includes: Acquire the audio recordings of the disputed conversations, and perform speaker analysis on the audio recordings to obtain the speaker analysis results; Based on the speaker analysis results, the disputed communication voice is segmented to obtain the speaker's voice, and the speaker's voice is then subjected to speech recognition to obtain the disputed speaker text. The disputed speakers' text is input into a pre-trained large model for mediation analysis to obtain mediation suggestions and risk levels. Based on the mediation suggestions, mediation prompts are sent to the mediators. If the risk level is greater than the level threshold, a dispute risk control alarm will be sent to the mediator. Speaker analysis was performed on the disputed communication audio to obtain speaker analysis results, including: The disputed conversation is segmented to obtain the segmented conversation, and the Mel frequency cepstral coefficients, fundamental frequency, and formant frequency of the segmented conversation are obtained. The Mel frequency cepstral coefficients, the fundamental frequency, and the formant frequency are combined to obtain segmented speech features, and the segmented speech features are vector-transformed to obtain segmented feature vectors. The communication segmented speech is classified according to the segmentation feature vector to obtain a segmented speech set, and the facial images of the disputants corresponding to the disputed communication speech are obtained. Based on the facial image, a set of pronunciation times is determined, and based on the set of pronunciation times, the speech correspondence between the person involved in the dispute and the segmented speech set is determined, thus obtaining the speaker analysis result.

2. The large-model-based assisted mediation method as described in claim 1, characterized in that, Before inputting the disputed speakers' text into the pre-trained large model for mediation analysis, the following steps are also included: Obtain a sample of the speakers in the dispute, and input the sample of the speakers in the dispute into the large model for word segmentation to obtain sample word segmentation; The sample words are segmented and feature extraction is performed to obtain context sample features, and the context sample features are decoded to obtain decoded sample features; Based on the features of the decoded samples, a mediation prediction is performed to obtain sample prediction suggestions and sample prediction levels, and the model loss is determined based on the sample prediction suggestions and sample prediction levels. The parameters of the large model are updated based on the model loss until the large model converges, thus obtaining the pre-trained large model.

3. The large-model-based assisted mediation method as described in claim 2, characterized in that, The sample word segmentation is used to extract features to obtain context sample features, including: The sample words are segmented and word embedding is performed to obtain word vectors, and the word vectors are then processed using a self-attention mechanism to obtain self-attention vectors; The self-attention vector is calculated using a multi-head attention mechanism to obtain a multi-head attention vector, and the multi-head attention vector is then normalized to obtain normalized features. The normalized features are subjected to pre-feedback processing to obtain pre-feedback features, and the pre-feedback features are subjected to residual processing to obtain the context sample features.

4. The large-model-based assisted mediation method as described in claim 1, characterized in that, The segmented speech is classified according to the segmentation feature vector to obtain a set of segmented speech, including: Calculate the similarity between adjacent segmentation feature vectors to obtain the first vector similarity; If the similarity of the first vector is greater than the first similarity threshold, then the speech segments corresponding to the segmentation feature vectors are merged to obtain the segmented and merged speech. Calculate the average vector of the segmentation feature vectors in the segmented and merged speech to obtain the segmentation average vector, and calculate the similarity between different segmentation average vectors to obtain the second vector similarity; If the second vector similarity is greater than the second similarity threshold, then the different segmented and merged speech corresponding to the second vector similarity are divided into sets to obtain the segmented speech set.

5. The large-model-based assisted mediation method as described in claim 1, characterized in that, The set of pronunciation times is determined based on the facial image, including: The face image is subjected to feature point recognition to obtain face feature points, and the lip image is determined based on the face feature points; Obtain the opening distance between the upper and lower lips in the lip image, and calculate the distance difference between the opening distances of adjacent lip images; If the distance difference is greater than the difference threshold, the adjacent lip image corresponding to the distance difference is determined as the speaking image, and the image acquisition time of the speaking image is obtained to obtain the pronunciation acquisition time; The sound acquisition time is stored in a preset set to obtain the sound acquisition time set.

6. The large-model-based assisted mediation method as described in claim 1, characterized in that, After sending a mediation prompt to the mediator based on the aforementioned mediation recommendations, the process also includes: Obtain mediation feedback on the mediation recommendation, and generate parameter fine-tuning samples based on the mediation feedback, the dispute speaker's text, and the mediation recommendation; The parameters of the pre-trained large model are adjusted by fine-tuning the samples based on the parameters.

7. A large-scale model-based assisted mediation system, characterized in that, The system includes: The speaker analysis module is used to acquire dispute communication voice recordings and perform speaker analysis on the dispute communication voice recordings to obtain speaker analysis results; The speech recognition module is used to segment the disputed communication speech according to the speaker analysis results to obtain the speaker's speech, and to perform speech recognition on the speaker's speech to obtain the disputed speaker's text; The mediation prompt module is used to input the text of the dispute speakers into a pre-trained large model for mediation analysis, obtain mediation suggestions and risk levels, and send mediation prompts to the mediators based on the mediation suggestions; if the risk level is greater than the level threshold, a dispute risk control alarm is sent to the mediators. The speaker analysis module is also used to: perform speech segmentation on the disputed communication speech to obtain the communication segmented speech, and obtain the Mel frequency cepstral coefficients, fundamental frequency and formant frequency of the communication segmented speech; The Mel frequency cepstral coefficients, the fundamental frequency, and the formant frequency are combined to obtain segmented speech features, and the segmented speech features are vector-transformed to obtain segmented feature vectors. The communication segmented speech is classified according to the segmentation feature vector to obtain a segmented speech set, and the facial images of the disputants corresponding to the disputed communication speech are obtained. Based on the facial image, a set of pronunciation times is determined, and based on the set of pronunciation times, the speech correspondence between the person involved in the dispute and the segmented speech set is determined, thus obtaining the speaker analysis result.

8. The large-model-based assisted mediation system as described in claim 7, characterized in that, The mediation prompt module is also used for: Obtain a sample of the speakers in the dispute, and input the sample of the speakers in the dispute into the large model for word segmentation to obtain sample word segmentation; The sample words are segmented and feature extraction is performed to obtain context sample features, and the context sample features are decoded to obtain decoded sample features; Based on the features of the decoded samples, a mediation prediction is performed to obtain sample prediction suggestions and sample prediction levels, and the model loss is determined based on the sample prediction suggestions and sample prediction levels. The parameters of the large model are updated based on the model loss until the large model converges, thus obtaining the pre-trained large model.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images